Anthropic today rolled out Computer Use, the Browser Tool, Skills API, and Files API to general availability. They are available for use directly, no longer marked as beta. This means Claude can now see the screen on its own, move the mouse, press keys, and operate software that does not provide an API for integration.

Let's start with Computer Use. It gives Claude the ability to capture screenshots, and control mouse and keyboard to work in a virtual desktop. The most practical update this time is that Claude can now execute multiple consecutive actions in a single turn. Previously, each model call could only complete one operation. According to the official blog, tasks can now be completed with fewer calls and in less time. Community users estimate that multi-action execution reduces the number of round trips by 20-40%. Fewer round trips mean faster task completion and lower costs.

The actual highlight worth paying attention to is the newly added Browser Tool. Previous browser automation relied on pixel coordinates—if the page layout changed even slightly, the click would end up in the wrong place. The new tool exposes page structure to Claude, letting it operate based on element references and coordinates. No matter where a button moves on the page, Claude can still find it. This makes it far more reliable for webpage-based tasks than generic Computer Use.

Skills API and Files API are built for agent development. Skills API lets teams package and upload their internal workflows with versioning, so you can pin requests to a specific version. Files API lets you upload a file once and reference it later via file_id, eliminating the need for repeated uploads. Today's update also added file expiration settings, increased rate limits to 500 RPM, and allocates 1TB of storage per organization. These features form the foundational infrastructure for building managed agents.

Anthropic's blog shares two sets of customer performance data. Asteroid reports that their claims processing workflow was reduced from 32 minutes to 13 minutes, cut per-task costs by roughly 30%, and reached a 100% completion rate—all without modifying their prompts. Box used Skills API to package banks' credit methodology and memo templates, letting Box AI agents generate credit memos directly. All these figures come from the official blog, and actual performance requires independent testing to verify.

Some commenters online note that most agent demos fail first due to cost, not capability—because every click requires processing an extra screenshot. They point out that versioned skills are an underrated feature, because this is what separates a proper agent from just a prompt you tweak over and over. Others joked: APIs were originally invented so we wouldn't have to click buttons, and now agents are clicking them for us. Despite the joke, this is exactly the positioning of Computer Use: legacy systems that never got an API can finally be automated.